MEASUREMENTS USING NOISE-DRIVEN NONLINEAR MICRO-RESONATORS By Pavel M. Polunin A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Mechanical Engineering - Master of Science 2013
ABSTRACT MEASUREMENTS USING NOISE-DRIVEN NONLINEAR MICRO-RESONATORS By Pavel M. Polunin In this work we discuss new measurement applications of nonlinear micro-resonators that are subject to stochastic forcing in combination with primary periodic excitation. First, we describe how the noise-driven switching of a nonlinear oscillator subjected to harmonic excitation with bistable response can be used as a new type of sensor for measuring parameter changes in the system; we refer to this as the balanced dynamic bridge. For small noise we develop a predictive theory that describes the dynamical behavior of these systems on the oscillator time scale, as well as on the characteristic time scale of the switching. A general theory of activated escape in the presence of a Gaussian noise allows one to compute the switching rates between the two stable states, and the manner in which these rates are related to system parameters. We discuss the high sensitivity of the bridge with respect to changes of system parameters and derive expressions describing the precision of the method and the time required to perform experimental measurements to a given precision. In addition, we discuss application of the dynamical bridge as a sensor of non-gaussian noise in the presence of additional weak non-gaussian stochastic forcing. We show how non-gaussian noise can be detected and estimated by measuring the long-time occupation probability ratio of the bistable states of the system. We conclude our discussion with two examples: they describe non-gaussian noise estimation in a one-dimensional system and in the nonlinear Mathieu resonator.
ACKNOWLEDGMENTS I would like to express all my gratitude and respect to my research advisor Professor Steven W. Shaw. Apart from his professional guidance and supervision, Professor Shaw has been my excellent teacher and instructor in all issues that I ever had during the time I worked on this thesis. I am really glad that fortune has given me the chance to work with this outstanding person. Additionally, I would especially like to thank Professor Mark I. Dykman for very fruitful discussions and valuable comments on my work process. Thanks also to committee members who oversaw this work: Professor Brian Feeny and Professor Ranjan Mukherjee. I owe my inspiration to my former lab-mate Nick Miller for a great number of useful discussions and his example of successful junior researcher; I am sure it will play important role in my future life as a PhD student. Furthermore, I thank my fellow labmates for providing a friendly atmosphere in the lab that has made me feel welcome and comfortable. These include Scott Strachan, Brendan Vidmar, Venkat Ramakrishnan, Smruti Panigrahi, Abhisek Jain and Xing Xing. I also thank to my friends, Andrey Maslennikov and Oleksii Karpenko, for their support and making my work more enjoyable. Finally, I would like to thank the funding agencies that made this work possible. The National Science Foundation under grant CMMI-0900666 and the Defense Advanced Research Agency under the DARPA-MTO-DEFYS project. iii
TABLE OF CONTENTS List of Figures................................ v Chapter 1 Introduction............................... 1 Chapter 2 The Balanced Dynamical Bridge.................. 7 2.1 Theory of activated escape with Gaussian excitation.............. 8 2.2 The Duffing Resonator as a Balanced Dynamic Bridge............ 13 2.3 Sensitivity of the Dynamic Bridge........................ 23 2.4 Summary..................................... 31 Chapter 3 Non-Gaussian Noise Detection................... 33 3.1 Switching in the Presence of Non-Gaussian Noise............... 34 3.2 Detection Technique............................... 39 3.3 An Example: Shot Noise Measurement in 1D.................. 42 3.4 Noise in Parametrically Excited Resonators................... 47 3.5 Summary..................................... 60 Chapter 4 Conclusions............................... 62 Appendix......................................... 65 Bibliography....................................... 69 iv
LIST OF FIGURES Figure 1.1 Shift in natural frequency of a linear resonator, as a tool for estimating parameter changes. (For interpretation of the references to color in this and all other figures, the reader is referred to the electronic version of this thesis.).......................... 3 Figure 1.2 Nonlinear resonator frequency response for different amplitudes of excitation. Dashed lines represent responses that are below the noise level while responses in solid lines exceed the noise level, and exhibit nonlinearity. This demonstrates a case where a linear dynamic range does not exist............................... 4 Figure 2.1 Solution to Eq. (2.27). The solid line represents the function g with Ω = 4.7 while the dashed line represents β = 8.5. The fixed points correspond to the intersection points, which provide the steady state values of u 2. Generally the outer two solutions are dynamically stable while the central solution is unstable............... 17 Figure 2.2 The phase portrait of a Duffing resonator depicted in the u phase plane Eqs. (2.26a) and (2.26b). The solid lines represent the unstable manifolds of the saddle, which originate at the saddle and terminate at the stable fixed points, while the dashed line designates the stable manifolds of the saddle, representing separatrices between the two basins of attraction. The parameter values are Ω = 4.7 and β = 8.5. 18 Figure 2.3 The (Ω, β) parameter space of the Duffing resonator, depicting the bistable region. The solid branches enclose the region of bistability and represent the bifurcation curves given in Eq. (2.30)........ 19 Figure 2.4 The heteroclinic solution to Eqs. (2.31a) to (2.31d) projected onto (u r, u i ) plane. The solid lines represent noisy dynamics and the dashed lines are the noise-free saddle manifolds. The parameter values are Ω = 4.7 and β = 8.5....................... 20 v
Figure 2.5 The heteroclinic solution to Eqs. (2.31a) to (2.31d) projected onto (p r, p i ) plane. The parameter values are Ω = 4.7 and β = 8.5..... 21 Figure 2.6 The (Ω, β) parameter space for the Duffing resonator. The solid lines represent the boundaries of the bistable region while the dashed line depicts the balance curve......................... 22 Figure 2.7 The occupation probabilities as a functions of the action change, S. 27 Figure 2.8 The lower bound on the number of measurements required to obtain an accurate estimate of the parameter change using the balanced dynamical bridge............................. 28 Figure 2.9 The lower bound on the required total measurement time as a function of the shift in the occupation probability. The dimension of the vertical axis, T m r bp, represents the number of the switching events that would happen if the system stays at the balance point for the time T m.................................. 30 Figure 3.1 Schematic one-dimensional representation of the potential for a bistable nonlinear system, Eq. (3.1)........................ 35 Figure 3.2 The change in the probability ratio due to exposure the balanced dynamic bridge to a non-gaussian perturbation............ 40 Figure 3.3 The double-well potential Eq. (3.24); q 1 and q 2 denote positions of the stable fixed points, while q S refers to the saddle-point. Poisson pulses with positive area g change the switching rates, such that r 1 > r 2, therefore shifting the probability from q 1 towards q 2.......... 43 Figure 3.4 Change in the probability ratio due to addition of the zero mean shot noise, Eq. (3.59). Data are results from Monte-Carlo simulations with D = 0.04 and ν = 0.5. The inset shows the same quantities over a larger range, demonstrating the limitations of the approximate theory. 45 Figure 3.5 Dependence of the amplitude of steady-state oscillations on the normalized detuning parameter Ω. Parameter α is taken to be equal 17.................................... 50 Figure 3.6 Phase portrait of the nonlinear Mathieu resonator........... 52 vi
Figure 3.7 Figure 3.8 Figure 3.9 Figure A.1 Shaded area represents the bistable region of the nonlinear Mathieu resonator................................. 53 Solutions to Eq. (3.43) are represented by the intersection points of the solid and the dashed lines. The left-hand side of Eq. (3.44) as a function of u o 2 is drawn by the solid line, while β 2 is drawn by the dashed line. Parameters values: α = 17, Ω = 0, ψ = 30, β = 1... 54 The numerical solution to the four-dimensional system of equations corresponding to the nonlinear Mathieu resonator depicted as a projection onto: (a) the (u r, u i ) plane, and (b) the (p r, p i ) plane. The noise-free dynamics is exactly the same as shown in Fig. 3.6(b), and is shown in black dotted lines. Red and blue lines are the projections of the noisy switching trajectories from different stable states to the saddle................................... 58 Schematic representation of a quantum-mechanical system that has more than one possible trajectory between points A and B. Different paths have different probabilities of realization, and these paths can significantly deviate one from another.................. 66 vii
Chapter 1 Introduction It is well known that noise can have a significant effect on system response in many applications, and this is especially the case when the characteristic size of the system of interest is very small. Here we use the term noise to designate various types of stochastic processes that can occur in systems. These include small fluctuations of charge on capacitors in electric circuits, tiny mass or stiffness fluctuations in mechanical systems, thermal vibrations, and so on. In most systems these noise processes have a negative effect on system response, and consequently, engineers search for ways to eliminate them or reduce their effects. However, this task becomes more challenging for systems at micro- and nano- length scales. In fact, for many systems, these random processes are an essential part of system dynamics and must be taken into consideration in system analysis and design. In this work we discuss vibrating micro/nano-electro-mechanical systems (M/NEMS) that operate in a nonlinear regime and are subject to stochastic effects in addition to resonant excitation. These systems are relevant to a number of existing and potential applications, including the counting statistics of charge transport systems, and employment as 1
elements in small-scale frequency generators. Of particular interest are systems that exhibit nonlinear responses, for which we give a detailed analysis of their dynamics and show that noise appearing in these systems can be constructively employed in particular applications, specifically: (i) for the measurement of parameter changes in the system, which can be linked to the system environment for purposes of sensing, and (ii) for the detection of specific types of non-gaussian random processes. In this chapter we give a brief review of M/NEMS, discuss how noise and random processes are related to M/NEMS, and conclude the chapter by providing more specific motivation for the present work. Having more than 50 years of development and enhancement [13], M/NEMS have found their place in many areas of modern human life. These include applications in microelectronics, medicine, aerospace and automotive engineering, and so on. One classification of M/NEMS divides most existing systems into two major groups: sensors like inertial, pressure and flow sensors, to name a few, and devices used for signal processing, for example filters, oscillators, switches, and relays. In this study we focus our attention on micro-mechanical resonators, which have applications in both groups. It is well known that linear systems are much easier to understand and analyze than their nonlinear counterparts, cf.[16]. So, there is no surprise that researchers initially worked with simple linear models of resonators for different types of measurement and processing purposes. Essentially, the concept of M/NEMS measurements using linear resonance is the following: suppose we have a micro-resonator that is coupled to a environmental parameter of interest in such a way that changes of this parameter will affect the resonant frequency of the system, see Fig. 1.1. Then, by calibrating and measuring these shifts in the natural frequency of the resonator, we can determine the associated change of the parameter of interest. For 2
example, many mass-sensing devices and humidity sensors work on this principle [12, 26]. Response Ω 0 Figure 1.1: Shift in natural frequency of a linear resonator, as a tool for estimating parameter changes. (For interpretation of the references to color in this and all other figures, the reader is referred to the electronic version of this thesis.) However, new fabrication techniques are making it possible to fabricate M/NEMS of smaller and smaller dimensions and, as mentioned previously, noise effects can come into play, and are often an essential component of the system response. Noise leads to corruption of the response signal, and, as a consequence, reduces the bandwidth of the system. In some cases, the ambient noise signal of a device can even overcome the signal containing measurement information, and the device loses its usefulness. The way around this obstacle is to increase the amplitude of system excitation in order to reach a desired signal-to-noise ratio. However, increasing the system input signal inevitably leads to a transition to a nonlinear response regime, where linear model approximation fail, see Fig. 1.2. In this situation one must consider the effects of nonlinearities and noise in the response, that is, a linear dynamic range does not exist. Nonlinear behavior of M/NEMS resonators involves many interesting aspects, including so-called bistability, that is, the property of the system to have multiple stable operating 3
Response Noise level Figure 1.2: Nonlinear resonator frequency response for different amplitudes of excitation. Dashed lines represent responses that are below the noise level while responses in solid lines exceed the noise level, and exhibit nonlinearity. This demonstrates a case where a linear dynamic range does not exist. regimes [21]. It turns out that noise in these cases can cause switching of the micro-resonator response between these stable operating regimes, cf. [2, 3, 8, 9, 20]. The rates of these transitions depend exponentially on the height of the potential barrier between stable states and on the noise strength, according to the Arrhenius law [1, 25], [ r exp R ], (1.1) D where R is the height of potential barrier and D is intensity of the noise. In the weak noise limit, i.e., when D is small, even small changes in the activation energy can produce significant changes in switching rates, which are reflected in system behavior. This sensitivity is a primary motivation for the present investigation. Specifically, we are interested in the possibility of using switching in noisy bistable systems for purposes of sensing. The significant part of the background and motivation to this work lies in Ph.D. thesis of N.J.Miller, [20]. In this work we consider two topics that are closely connected and represent novel alternatives to contemporary measurement techniques. First is the application of noise-driven 4
bistable systems for parametric sensing, which we refer to as the balanced dynamical bridge. In Chapter 2 we consider the bridge application for the measurement of changes in system parameters due to environmental changes. The theory of activated escape forms the basis of the analysis of the bridge, and this is described first. We then consider a Duffing resonator as a model for our subsequent discussion and analyze its switching dynamics in detail. We derive and analyze a model describing transitions of the system between its stable states and investigate it numerically. Finally, we turn to the switching dynamics on a long time scale and derive expressions that describe the measurement precision of the bridge, as well as the time needed for measurements. In Chapter 3 we examine the second application of M/NEMS, specifically, use of the balanced dynamical bridge as a detector of non-gaussian noise. The presence of an additional non-gaussian modulation to a bistable system switching under Gaussian noise requires a substantial change in the theory of the activated escape, and we address this modification at the beginning of Chapter 3. We find that this non-gaussian forcing is projected onto the so-called susceptibility of the system on switching trajectories in the phase space. This projection affects the switching rates of the resonator in each direction differently, thus providing a shift of the population ratio. Using these results we derive expressions that allow us to determine the statistical properties of the non-gaussian noise from changes in transition rates. Two examples of non-gaussian detection are presented. The first example considers a simple one-dimensional system where analytical predictions can be completed in closed form, and the second example deals with a nonlinear Mathieu resonator, which is substantially more complex and requires more numerical calculations. The results from these examples allow us to draw some general conclusions about the effectiveness of using 5
switching in noisy bistable M/NEMS for measurement and detection purposes. 6
Chapter 2 The Balanced Dynamical Bridge In this chapter we discuss a new measurement paradigm which we refer to as balanced dynamic bridge. This sensing application can be realized in bistable systems that are driven by a white Gaussian noise such that the system randomly switches between the two stable states. Measurement is achieved by relating the ratio of the occupation probabilities of the two states to system properties that are affected by the system environment. It is shown that this measurement approach is most sensitive when one operates near a point where the occupation probabilities are equal. It is also shown that this method has high resolution, but longer measurement time, when compared to more standard measurement techniques. In Section 2.1 we describe a general model and theory for switching dynamics of bistable systems in the presence of white Gaussian noise. In Section 2.2 we consider the specific example of a resonantly driven Duffing oscillator. In Section 2.3 we describe a general theory for measurement sensitivity and measurement time for the balanced bridge and show results derived from the Duffing example. The Chapter is closed with a Summary in Section 2.4. 7
2.1 Theory of activated escape with Gaussian excitation In this section we review a general theory of noise-activated escape in the presence of a Gaussian noise. Originated by Kramer in his famous work [17], this theory has been extended significantly since that time and has found its place in the many applications. We start with a general system those dynamics can be expressed in the state variable form as follows, u = K(u) + ˆf(t), (2.1) where u R n is the state vector of the system of interest, K is the deterministic nonlinear vector field, and ˆf R n is a stationary Gaussian noise vector with characteristic strength D. It is assumed that form of K gives rise for the system to experience at least two stable steadystates, and that all components of the noise vector, ˆf, are delta-correlated and independent from each other. For further convenience we transform the system of interest to a discrete form, u i+1 u i t = K(u i ) + ˆf i, (2.2) using the following time-slicing t = t i+1 t i, i = 1, 2,..., N. (2.3) Since ˆf is the vector of the independent white Gaussian noise processes, its auto-correlation function is simply ˆf i ˆf T j = D t δ iji, (2.4) 8
where δ ij denotes the Kronecker delta function and I is the identity matrix. Using these introductory concepts we can formulate the main results of this section, namely we can compute the probability for the system to make a transition from a region near a stable fixed point in the configuration space to another region, near another fixed point, typically a saddle point, during a given time interval. Since the noise is considered to be small, such transitions rare, and in order to be realized they require quite specific time evolutions of the stochastic driving. In the case when system dynamics is represented in the form of Eq. (2.2), this time evolution of ˆf is described in terms of the { ˆf i }. Because of the δ-form of the autocorrelation function for ˆf, we have N independent noise vectors ˆf i. Each of these vectors, in turn, has Gaussian probability distribution given by, p( ˆf i ) = [ 1 exp ˆf i 2 ] 2πσi 2 2σi 2, (2.5) where σ i = D/ t is constant for all elements ˆf i. During the time evolution of the system, N noise bursts occur in a series and we are interested in particular sequence of these bursts which will force the system to move between points of interest in the configurations space. Following Jacobs [15], the joint probability distribution of these independent processes is the product of individual probability densities for each noise vector p({ ˆf}) = p( ˆf 1, ˆf 2,..., ˆf N N ) = p( ˆf i ). (2.6) i=1 Applying Eq. (2.5) to the expression for the joint probability distribution, we find [ p({ ˆf}) = C N exp 1 2D 9 N i=1 ] ˆf i 2 t, (2.7)
where C N is a coefficient that absorbs all exponential prefactors. Now it is possible to calculate the probability that the system will come to point u 2 at time t 2 being near point u 1 at time t 1. This can be done by using the standard relationship between the cumulative probability function and its density distribution, yielding [15] P (1 2) = P (u 2, t 2 ; u 1, t 1 ) =... p({ ˆf})δ(u[u 1, { ˆf}) u 2 ]d ˆf 1 d ˆf 2... d ˆf N, (2.8) where we have introduced δ-function term reflecting the fact that the sequence {f} must bring the system from point u 1 to point u 2. By changing the variables of integration from d ˆf i to du i, and using Eqs. (2.2) and (2.7), we obtain the following expression for the total probability, P (1 2), P (1 2) = C N u 2 u 1 [ det(j) exp 1 2D N i=1 u i+1 u i t 2 ] K(u i ) t δ(u u 2 )Du, (2.9) where J is the Jacobian of the [ ˆf u] transformation and Du designates the path integral between points u 1 and u 2. Appendix A describes more details of this development. Since the Jacobian J describes the transformation between different system parameters, it remains fixed for a given system and, as a result, it is independent of the path the system follows between given points in the phase space. So, we can also include its determinant in the integral prefactor, and by taking the limit as t 0, we come to the final form for the transition probability expressed in terms of the system dynamics u 2 P (1 2) = C u 1 [ exp 1 t ] 2 u K(u) 2 dt Du. (2.10) 2D t 1 10
It is useful to note here that there are an infinite number of trajectories that the system can follow as it undergoes the transition between two given points in the state space. However, the probability of the realization of a particular trajectory depends strongly on the path, and so there exists a probability distribution for the possible trajectories. Then, the usual way to calculate the total probability for for the system of interest to switch from one state to another is to integrate this probability density distribution over the whole range of possible trajectories. That is exactly the role of the path integral in Eq. (2.10); it simply integrates the probability distribution over all possible time evolutions of the system between two given points. The probability distribution discussed above possesses one remarkable property which significantly facilitates the calculation of the total transition probability. According to previous work of Dykman and collaborators, cf.[9, 2, 3], this probability distribution is sharply peaked about so-called most probable path. Using this property, we can approximate the total transition probability by the expression in the integrand of Eq. (2.10) evaluated along the most probable path, as follows, [ P (1 2) C exp 1 t 2 u K(u) dt] 2. (2.11) 2D t 1 MP P Hereafter we assume that we only work with the most probable path. According to the path integral formulation, the transition probability P (1 2) depends exponentially on the action, calculated along the most probable path, namely, ( P (1 2) exp S ) 1 2 D (2.12) 11
Then, by comparison Eqs. (2.11) and (2.12) one can see that the action S 1 2 can be expressed in the following form S 1 2 = 1 t 2 u K(u) 2 dt. (2.13) 2 t 1 On the other hand, it is well known from the classical mechanics [11] that mechanical action is related with the Lagrangian of the system as t 2 S = L(q, q, t)dt, (2.14) t 1 So, the Lagrangian for the general system, Eq. (2.1), reads L(u, u) = 1 2 u K(u) 2 (2.15) along with the generalized momenta defined as p = L u = u K(u), (2.16) which is exactly the vector of white Gaussian noise processes defined in Eq. (2.1). Now, we only have to specify the equations of motion for the generalized momenta in order to get a complete description of the system dynamics. It can be done in the following way, ṗ = H u = L u = K u p, (2.17) where H is the associated Hamiltonian of the system. Eqs. (2.1) and (2.17) are the full set of equations of motion that completely describe the system dynamics in the extended R 2n state 12
space, which contains the noise-free dynamics and the realizations of noise. The solution to these equations describing the heteroclinic trajectory between two stable states represents the most probable path for the system in this space. 2.2 The Duffing Resonator as a Balanced Dynamic Bridge In this section we discuss the harmonically excited Duffing resonator as an example of a nonlinear system that exhibits bistability under certain conditions, and its operation as a dynamic bridge when subjected to additive noise. We start from the model of the Duffing resonator, subject to both periodic and stochastic excitation, q + 2Γ q + ω 2 0 q + γq3 = h cos ωt + Df(t), (2.18) where q is the resonator coordinate, ω 0 is its natural frequency, Γ is its damping coefficient, γ is the nonlinearity parameter, h is the amplitude of the harmonic forcing, ω is the driving frequency, and Df(t) is white Gaussian noise of strength D. It is assumed that the Gaussian noise has zero mean, f(t) = 0, and is delta-correlated, f(t)f(t ) = 2Dδ(t t ). We also assume that ω ω 0 ω, ω 0, that is, the system is driven near resonance and is lightly damped, Γ ω 0. In order to simplify the analysis, it is convenient to switch from coordinate q to a slowly-varying complex amplitude u in a rotating frame, achieved by the well known van der Pol transformation, q(t) = 2ωΓ 3 γ ueiωt + c.c., ue iωt + c.c. = 0. (2.19a) (2.19b) 13
Using this transformation, we can obtain the following expressions for the first and second derivatives of the primary variable q, namely q(t) = 2ωΓ 3 γ iω(ueiωt u e iωt ), q(t) = ω 2 2ωΓ 3 γ (ueiωt + u e iωt ) + 2iω 2ωΓ 3 γ ueiωt. (2.20a) (2.20b) Substitution of q and its derivatives as functions of u, u, u, and u into Eq. (2.18), elimination of fast oscillating terms of the deterministic part of the equation, and rescaling of time and parameters yields the following complex first order differential equation for the slow-varying amplitude u, u = (1 + iω)u + i sgn(γ) u 2 u i β i ˆf(τ), (2.21) where ( ) is understood to be the derivative with respect to the rescaled time τ, defined below. Additionally, in order to make the equation fully non-dimensional, we have introduced the following quantities, t = Γτ, (2.22a) Ω = ω2 ω 2 0 2Γω, (2.22b) β = 3 γ h2 32ω 3 Γ 3, 3 γ ˆf(τ) = 8ω 3 Γ 3 e iωτ/γ f(τ/γ). (2.22c) (2.22d) Since the spectrum of the white noise f(t) has constant magnitude over infinite frequency range, and because we are interested in switching dynamics of the system, we have not averaged the noise term, and thus we retain the complex exponential factor in Eq. (2.22d). 14
Instead of trying to extract its mean value and neglect fast oscillating components, we use Euler s formula for the complex exponent to separate real, ˆf r (τ), and imaginary, ˆf i (τ), parts of the complex noise function, resulting in ˆf(τ) = 3 γ 8ω 3 Γ 3 ( ˆf r (τ) i ˆf i (τ)), (2.23) where ˆf r (τ) = ˆf(τ) cos(ωt) and ˆf i (τ) = ˆf(τ) sin(ωt). These projections of the original white Gaussian noise onto two orthogonal quadratures makes them approximately independent white noise processes with following properties [9], ˆf r (τ) = ˆf i (τ) = 0, (2.24a) ˆf r (τ) ˆf r (τ ) = ˆf i (τ) ˆf i (τ ) = ˆDδ(τ τ ), (2.24b) ˆf r (τ) ˆf i (τ ) = 0, (2.24c) where ˆD is the nondimensional noise intensity of each quadrature, defined as, ˆD = 3 γ D 8ω 3 Γ 3. (2.25) Equation (2.21) describes the time evolution of the complex slowly-varying amplitude u. In the absence of the stochastic excitation, it is possible to find stationary solutions of the resonator by setting u = 0, cf.[16]. In order to determine these stationary amplitudes, we 15
first separate real, u r, and imaginary, u i, parts of u and get the following dynamical system, u r = u r + Ωu i sgn(γ)(u 2 r + u 2 i )u i, u i = u i Ωu r + sgn(γ)(u 2 r + u 2 i )u r β. (2.26a) (2.26b) Now, by setting u r = u i = 0 and taking a convenient linear combination of the resulting equations we obtain, g( u 2, Ω) = u 6 2 sgn(γ)ω u 4 + (1 + Ω 2 ) u 2 = β, (2.27) where u is the magnitude of u and function g is introduced for convenience. Equation (2.27) is a cubic algebraic equation in terms of u 2 which, in general, provides three stationary solutions to the system of Eqs. (2.26a) and (2.26b). Since we now work with u 2, we are only interested in the real roots of Eq. (2.27). The number of real roots of any cubic equation depends on the values of the coefficients of the polynomial, which in this case depend on the values of Ω and β. Fig. 2.1 shows a sample of the cubic function g versus u 2 and a sample level line of β. There are two qualitatively different types of system dynamics. Depending on the values of Ω and β, the system of interest generically has either one or three fixed points in the u phase plane. In the latter case, two fixed points are stable while the third one is an unstable saddle point[16]. A system that exhibits this type of dynamics is called bistable. The typical phase portrait of a bistable Duffing resonator is depicted in Fig. 2.2, where the bistability of the system and the attendant domains of attraction are clearly shown. The basins of attraction of the two stable fixed points are delineated by separatrices, that is, the stable 16
15 g u 2, Β 10 5 0 0 1 2 3 4 5 6 Figure 2.1: Solution to Eq. (2.27). The solid line represents the function g with Ω = 4.7 while the dashed line represents β = 8.5. The fixed points correspond to the intersection points, which provide the steady state values of u 2. Generally the outer two solutions are dynamically stable while the central solution is unstable. u 2 manifolds of the saddle point. Since the concept of the balanced dynamical bridge is based on the switching dynamics between the system s stable states, it is reasonable to determine the region in the (Ω, β) parameter space where the Duffing resonator has two stable solutions. In order to find this region of bistability, we can look at Fig. 2.1 and note that for a fixed value of Ω, which determines the form of the cubic function g, there are two values of β where a stable and an unstable solution merge. These are saddle-node bifurcations [14], and it is clear that the derivative of the cubic function g with respect to u 2 is equal to zero at these bifurcation points, i.e., 3 u 4 4 sgn(γ)ω u 2 + (1 + Ω 2 ) = 0. (2.28) Equation (2.28) is quadratic in terms of u 2, and we can solve it obtaining the following 17
u i 1 0.5 0-0.5-1 -1.5-2 -2.5-2 -1 0 1 2 u r Figure 2.2: The phase portrait of a Duffing resonator depicted in the u phase plane Eqs. (2.26a) and (2.26b). The solid lines represent the unstable manifolds of the saddle, which originate at the saddle and terminate at the stable fixed points, while the dashed line designates the stable manifolds of the saddle, representing separatrices between the two basins of attraction. The parameter values are Ω = 4.7 and β = 8.5. values of u 2 at the bifurcation points as a function of Ω, u 2 B = 1 3 {2 sgn(γ)ω ± Ω 2 3}. (2.29) In order to obtain the bifurcation curves in (Ω, β) parameter space, one substitutes the values of u 2 B into Eq. (2.27), which yields the saddle-node parameter condition for β, β B = 2 27[ Ω(Ω 2 + 9) sgn(γ) ± (Ω 2 3) 3/2], Ω 2 > 3. (2.30) The bifurcation curves obtained from Eq. (2.30) are depicted in a convenient form in Fig. 2.3. 18
0.3 Β 3 0.2 0.1 0.0 0.0 0.1 0.2 0.3 4 Figure 2.3: The (Ω, β) parameter space of the Duffing resonator, depicting the bistable region. The solid branches enclose the region of bistability and represent the bifurcation curves given in Eq. (2.30). By defining the bistable region in the (Ω, β) parameter space, we have completely described the deterministic part of the system dynamics. In order to understand the switching dynamics of the resonator, we use the results of Section 2.1, namely, Eqs. (2.1) and (2.17) capturing the noise dynamics. Thus, the equations of motion governing the optimal switching trajectories are four-dimensional and have the following form, u r = u r + Ωu i sgn(γ)(u 2 r + u 2 i )u i + p r, u i = u i Ωu r + sgn(γ)(u 2 r + u 2 i )u r β + p i, ṗ r = [1 + 2 sgn(γ)u r u i ]p r + [(u 2 r + 3u 2 i ) sgn(γ) Ω]p i, ṗ i = [Ω sgn(γ)(u 2 i + 3u2 r)]p r + [1 2 sgn(γ)u r u i ]p i. (2.31a) (2.31b) (2.31c) (2.31d) Solutions to this system of equations represent the most probable path the system follows in the extended state space, where the u s are the system states and the p s represent noise 19
1 0.5 0-0.5 u i -1-1.5-2 -2.5-3 -2-1.5-1 -0.5 0 0.5 1 1.5 2 u r Figure 2.4: The heteroclinic solution to Eqs. (2.31a) to (2.31d) projected onto (u r, u i ) plane. The solid lines represent noisy dynamics and the dashed lines are the noise-free saddle manifolds. The parameter values are Ω = 4.7 and β = 8.5. quadratures. It is important to note that the noise-free dynamics are invariant, that is, if one starts with zero initial conditions for the p s, they remain zero for all time, and the system simply follows the deterministic equations of motion. Thus, fixed points of the deterministic system are also fixed points of the full system. However, the stability of these points can be different; in fact, fixed points that are stable in the deterministic system become saddle points in the full system, and heteroclinic trajectories that connect deterministic fixed points to one another, via a noisy trajectory, are an essential feature of the switching dynamics. The two optimal switching trajectories that go between the two deterministically stable foci, one each way, each consist of two parts. The first part is induced by the noise and is 20
1.5 1 0.5 p i 0-0.5-1 -1.5-0.6-0.4-0.2 0 0.2 0.4 0.6 0.8 1 p r Figure 2.5: The heteroclinic solution to Eqs. (2.31a) to (2.31d) projected onto (p r, p i ) plane. The parameter values are Ω = 4.7 and β = 8.5. represented by the solution of Eqs. (2.31a) to (2.31d) that goes from a deterministically stable focus to the deterministic saddle; such a trajectory starts and ends with the p s both zero, but must extend into the noise states. The second half of a switching trajectory is essentially noise-free and goes along the corresponding unstable manifold of the saddle, asymptotically approaching the region of other deterministically stable focus. Of course, once the solution gets to the saddle point, there is a 50/50 chance of it going towards either deterministically stable focus. Thus, the two complete switching trajectories are each composed of the joining of two heteroclinic solutions, the first of which involves all four states, while the second is essentially deterministic. The challenge to determining these trajectories is that of finding saddle-to-saddle heteroclinic solutions of a four-dimensional system, since the deterministic part is simple to find. By using a shooting method, these heteroclinic solutions to the four 21
dimensional equations can be obtained, a sample of which is depicted in Figs. 2.4 and 2.5, depicted by two dimensional projections. The results shown in Figs. 2.4 and 2.5 were obtained for a special case when the action calculated along the two switching trajectories has the same value in both directions, which requires a constraint on parameters Ω and β. In this case we call the system a balanced dynamic bridge, which is especially useful for the measurement purposes, as described below. Additional numerical calculations show that there exists a so-called balance curve, that is, a locus of points in the (Ω, β) parameter space, where the system is in the balanced bridge configuration. This curve is depicted as the dashed line in Fig. 2.6. 0.25 0.20 Β 3 0.15 0.10 0.05 0.00 0.00 0.05 0.10 0.15 0.20 0.25 0.30 Figure 2.6: The (Ω, β) parameter space for the Duffing resonator. The solid lines represent the boundaries of the bistable region while the dashed line depicts the balance curve. The results presented above set the stage for using the Duffing resonator with additive Gaussian noise as a balanced bridge. In the next section we discuss some general features of the bridge as a measurement tool. 22
2.3 Sensitivity of the Dynamic Bridge In this section we discuss the quantitative aspects of the balanced dynamical bridge as a measurement tool. The idea is to use the occupation probability ratio of the two stable states, and its dependence on system and input parameters, as the source of information. Since this ratio depends on the switching rate, we can use the results derived above to develop a theoretical basis for such a sensor. Specifically, we investigate the sensitivity properties of the bridge and the time required to perform measurements to a given level of precision. Section 2.2 describes the dynamics of the resonator on the time scale of the order of one switching event. In order to analyze the performance of this measurement techniques, we need to consider a longer time scale since the probability ratio is a statistical quantity of the system accumulated over many switching events. Considering the transitions that the system makes between the two stable states on a long time scale, it is convenient to use a master equation governing the dynamics of a two-level system, given by [23], P 1 = r 1 P 1 + r 2 P 2, P 2 = r 1 P 1 r 2 P 2, (2.32a) (2.32b) where P i is the probability to find the system in i th state, also known as the occupation probability, and r i is the switching rate out of the i th state, for i = 1, 2. Since there are just two states that system can occupy, the occupation probabilities must satisfy the normalization condition, namely, P 1 + P 2 = 1. (2.33) Hereafter, we focus our attention only on P 1, assuming that P 2 and its dynamics can be found 23
using this condition and the master equation. The time-dependent solution to Eqs. (2.32a) and (2.32b) can be obtained by using standard techniques [7, 16] and reads, P 1 (t) = r ) 2 (1 e (r 1 +r 2 )t + P r 1 + r 1 (0)e (r 1 +r 2 )t. (2.34) 2 From this expression it is clear that the characteristic time constant of the bridge is quite large, namely t r (r 1 + r 2 ) 1. The concept of the dynamic bridge as a sensor is to track changes in occupation probabilities that vary as a parameter of interest changes. In order to measure the occupation probability, P 1, we have to have some quantity x in the system that takes on the value X 1 when the resonator is in the first state and X 2 when the resonator is in the second state. Suppose we take N measurements of the system states. The time between successive measurements has to be long enough for the system to relax; particularly, it has to be on the order of t r, in order to statistically capture the switching dynamics. Taking this into account, we can conclude one requires a time on the order of N t r, with N sufficiently large (as described below) in order to perform the measurement with sufficient accuracy. Having x as the test quantity, we can write the following expression for the estimator of P 1 [19], ˆP 1 = 1 N(X 1 X 2 ) N (x i X 2 ). (2.35) i=1 The mean and the variance of the estimator ˆP 1 are then ˆP 1 = P 1, (2.36a) ( ˆP 1 P 1 ) 2 = P 1 P 2 1 N, (2.36b) 24
where P 1 = r 2 r 1 + r 2 (2.37) is the steady-state value of P 1. Equation (2.36b) shows that the uncertainty of the probability measurement is proportional to N 1/2, which proves that the accuracy of the measurement technique is higher when more measurement experiments are performed. Now, suppose that some parameter λ of the resonator changes by an amount λ. Expansion of P1 in a Taylor series yields that the system sensitivity is described by d P 1 /dλ. In the following calculation we assume that the prefactor of the switching rate depends on λ weakly, and we neglect this dependence in the weak noise limit. As a result, the corresponding change in the occupation probability is caused by the change of the activation energy in both directions and is given by, d P [ 1 dλ = r 1 r 2 (S1 S 2 ) D(r 1 + r 2 ) 2 S 1 S 2 λ D ] D, (2.38) λ where S i is the activation barrier that system has to overcome in order to escape from i th basin of attraction, i = 1, 2. This expression provides the link between the general analysis of the switching dynamics and the measurement sensitivity. Equation (2.38) shows that the operating point where the system exhibits the highest sensitivity with respect to the parameter change is located on the balance curve, as shown in Fig. 2.6 for the Duffing equation. In this case r 1 r 2 and S 1 S 2, and the sensitivity term simplifies to, d P 1 dλ 1 D (S 1 S 2 ), (2.39) λ which shows that the dynamic bridge can be extremely sensitive as the noise intensity D 25
decreases. After a sufficiently long time after a system parameter changes, the switching process relaxes and new occupation probabilities are established in the system. The corresponding shift in the occupation probability P 1 is proportional to the parameter change, namely, P 1 = d P 1 λ. (2.40) dλ In order to detect this change with the bridge, it is necessary that the uncertainty in the measurement process must be of the order of the probability shift that we want to measure, implying that, ( ˆP 1 P 1 ) 2 P 1. (2.41) This condition allows one to determine the number of measurements, N, required to detect the shift in the parameter value, assuming that the dynamic bridge is operating on the balance curve, so that P 1 = 1/2. When the parameter λ changes, it causes a corresponding change in the action, S, and using the fact that the switching rate depends exponentially on the path action [ r i exp S ] i 2D (2.42) and Eq. (2.37) one finds, P 1 + P 1 = e S 2D e S 2D + e S. (2.43) 2D In order to obtain an expression for P 1, we subtract the value of 1/2 of equilibrium prob- 26
ability P 1 from the RHS of Eq. (2.43) and simplify the result to P 1 = 1 ( ) S 2 tanh. (2.44) 2D Expressing S from Eq. (2.43), we have 1 P 2 0 P 1 Figure 2.7: The occupation probabilities as a functions of the action change, S. ( ) 1 + 2 P1 S = D ln 1 2 P. (2.45) 1 However, the estimated value of the action change, ˆ S, is slightly different from the S shown in Eq. (2.45). This deviation stems from the uncertainty in the occupation probability that arises from the finite number of measurements; see Eq. (2.36b). Up to the first order expansion, the estimation of the change in action reads, ( ) 1 + 2 Ŝ = D ln P1 4D 1 2 P 1 4 P ˆP 1 2 1( 1 P 1 ). (2.46) With these results in hand, we can formulate a condition for a target level of precision of the measurement process. The measurement is deemed sufficiently accurate if Ŝ S, which 27
is only possible if the second term of the expansion in Eq. (2.46) is much smaller than the first one, or ( ) ( ˆP 1 P 1 1 ) 2 4 P 1 2 ln ( 1 + 2 P1 1 2 P 1 ). (2.47) Additionally, following from Eq. (2.36b), the uncertainty in the occupation probability is proportional to (4N) 1/2 at the balance operating point, which gives one the opportunity to obtain the expression for the required number of measurements, given by, [ ( ) ( )] 1 N 2 4 P 1 2 1 + 2 2 P1 ln 1 2 P. (2.48) 1 The RHS of Eq. (2.48), with an equality in place of to indicate the boundary, is depicted 80 60 N 40 20 0 0.4 0.2 0.0 0.2 0.4 Figure 2.8: The lower bound on the number of measurements required to obtain an accurate estimate of the parameter change using the balanced dynamical bridge. 1 on Fig. 2.8. These curves, which are symmetric about zero, exhibit some interesting feature that demonstrate the special utility of the bridge, as discussed further below. 28
The attendant total measurement time, T m, is proportional to the product of the number of measurements N and the characteristic relaxation time of the resonator, t r, which can be expressed in the following way, ( t r (r 1 + r 2 ) 1 = rbp 1 e S 2D + e S ) 1 ( 2D = rbp 1 1 + 4 P 2 ) 1 1 1 4 P 1 2, (2.49) where r bp is the switching rate at the balance operating point. Then the expression for the lower bound on the required total measurement time reads as, T m 1 [( 1 r bp 4 3 ) ( )] 4 P 1 2 P 1 4 ln 2 1 + 2 1 P1 1 2 P, (2.50) 1 and this result is depicted in Fig. 2.9, which has the same general features as the required number of measurements N, which are discussed next. To sum up, we have described and proved that in general the measurement time is the longest time scale. In particular, it needs to be much longer than the period of the resonator oscillations, and much longer than its mean switching time. However, as Fig. 2.9 clearly shows, when the parameter λ changes only very slightly, it causes an infinitesimally small change in the occupation probabilities, due to the linear dependence (see Eq. (2.40)) and, consequently, the measurement process will take a very long time to detect this tiny deviation, because the variance of the estimated probability has to be less or equal to the infinitesimal change in P 1. As Eq. (2.36b) dictates, this variance is inversely proportional to N, and it only can converge to zero when N. This results in the unboundedness of N and T m as δ P 1 0. The situation is quite similar when P 1 approaches the value of 1/2, where clearly the occupation probabilities are nearly constant and therefore insensitive to 29
60 Tmrbp 40 20 0 0.4 0.2 0.0 0.2 0.4 Figure 2.9: The lower bound on the required total measurement time as a function of the shift in the occupation probability. The dimension of the vertical axis, T m r bp, represents the number of the switching events that would happen if the system stays at the balance point for the time T m. 1 parameter changes. Equation (2.36b) says that the occupation probability variance becomes very small as P 1 1. In this case the probability goes to its boundary value, and we call this process the probability saturation. So, the probability distribution for the estimated occupation probability will collapse to a sharp peak in the close vicinity of this boundary, and in the limit P 1 1, it approaches a delta-function, δ( P 1 1). Additionally, the information about the values of the occupation probabilities is based on the relative number of particular outcomes of the system, X 1 and X 2, and one would need a great number of experiments in order to detect the difference between the actual value of the occupation probability and its saturation value. In fact, in the limit P 1 1, the number of measurements becomes unbounded. These extremes are consistent with the trends shown in Figs. 2.8 and 2.9. However, between these limits there is a range of P 1, and thus a range of changes in λ, for 30
which the measurement time can be relatively short. This is the remarkable feature of the dynamic bridge as a measurement tool. 2.4 Summary The analysis presented in this chapter was originated by the idea to employ the noisy switching dynamics of bistable systems for purposes of sensing. The applicability of the dynamic bridge requires only that one couple the quantity of interest with the system and be able to measure shifts in the occupation probabilities. After presenting an introduction into the general theory of activated escape, we analyzed the Duffing resonator subject to the nearresonant periodic driving as an example of a bistable system, with additive white Gaussian noise to induce switching. In Section 2.2 we discussed the switching dynamics of the resonator on the time scale related to the relaxation time of the system. In the next section we quantified how bistable systems can be used as a measurement tool for detecting changes in system parameters, examining sensitivity and measurement time. We determined the sensitivity limits of the proposed paradigm and showed that the dynamic bridge can be extremely sensitive to shifts in system parameter values over a certain range. This allows one to detect and estimate small changes in the parameter of interest, although the method can require relatively long measurement times. However, we determined that, in an intermediate range of parameter shifts and the attendant occupation probabilities, the total measurement time is comparatively short. Additionally, the proposed technique allows one to track changes in large variety of system parameters, but only in those that change the natural frequency of the system. These results are quite promising since they provide confidence that this measurement paradigm can be competitive with other methods, 31